Baseline Prediction for TreeExplainer and XGBoost is not being computed correctly.

[mmschlk]

I think we might have broken something with XGBoost and TreeExplainer. It seems like the baseline values are not being computed correctly (it's always sitting at zero). Currently there is already a TODO about this in the code.

# TODO: for the current implementation this is correct for other trees this may vary
self.baseline_value = sum(
    [treeshapiq.empty_prediction for treeshapiq in self._treeshapiq_explainers]
)

I think this needs to be resolved either in the conversion from XGBoost models into TreeModels or in the Explainer. If it can be resolved in the conversion this would definetly be the preferred fix.

Code to reproduce this:

from xgboost import XGBRegressor
from sklearn.tree import DecisionTreeRegressor
from sklearn.ensemble import RandomForestRegressor
import shapiq
x_data, y_data = shapiq.datasets.load_california_housing(to_numpy=True)
x_train, x_test, y_train, y_test = train_test_split(x_data, y_data, test_size=0.2, random_state=42)
models = [
    XGBRegressor(random_state=42),
    DecisionTreeRegressor(max_depth=8, random_state=42),
    RandomForestRegressor(n_estimators=10, max_depth=8, random_state=42)
]
for model in models:
    print(f"{model.__class__.__name__}:")
    model.fit(x_train, y_train)
    explainer = shapiq.TreeExplainer(model=model, max_order=1, index="SV")
    sv = explainer.explain(x=x_test[0])
    baseline_value = round(sv.baseline_value, 3)
    sv = sv.get_n_order(min_order=1, order=1)
    sum_sv = round(sum(sv.values), 3)
    sum_sv_baseline = round(sum_sv + baseline_value, 3)
    y_pred = round(float(model.predict(x_test[0].reshape(1, -1))[0]), 3)
    print(f"Predicted Value: {y_pred}")
    print(f"Sum of SV + Baseline: {sum_sv_baseline}")
    print(f"Sum of SV: {sum_sv}")
    print(f"Baseline Value: {baseline_value}\n")

Produces:

XGBRegressor:
Predicted Value: 0.594
Sum of SV + Baseline: -1.477
Sum of SV: -1.477
Baseline Value: -0.0

DecisionTreeRegressor:
Predicted Value: 0.688
Sum of SV + Baseline: 0.689
Sum of SV: -1.383
Baseline Value: 2.072

RandomForestRegressor:
Predicted Value: 0.681
Sum of SV + Baseline: 0.681
Sum of SV: -1.391
Baseline Value: 2.072

For the decision tree and the random forest the predicted value and the sum of the SVs and the baseline are equal. This is to be expected. For the XGBoost model this is not the case.

[hbaniecki]

SIs don't sum to the model's prediction? See

import shapiq
# load data
X, y = shapiq.load_california_housing(to_numpy=True)
# train a model
from sklearn.ensemble import RandomForestRegressor
model = RandomForestRegressor()
model.fit(X, y)
# set up an explainer with k-SII interaction values up to order 4
explainer = shapiq.TabularExplainer(
    model=model,
    data=X,
    index="k-SII",
    max_order=2
)
# explain the model's prediction for the first sample
interaction_values = explainer.explain(X[0], budget=256)
# analyse interaction values
interaction_values.plot_force()

# vs
model.predict(X[[0]])